Eager Backend Dispatch

Cache each accepted backend capsule, then require an executable handler for that exact provider and device. Reports describe the selection before handler lookup or execution; they are not a record of completed native calls.

def Runtime.Autograd.Torch.Internal.EagerSession.selectedCapsule {α : Type} [TorchLean.Storage α] (s : EagerSession α) (op : NN.Backend.BackendOp) : IO NN.Backend.KernelCapsule

Select the accepted capsule for one operation.

The profile stays fixed for the session, so cache the first choice for each operation. executeSelected binds that contract to a handler for the same provider and device.

def Runtime.Autograd.Torch.Internal.EagerSession.executeSelected {α β : Type} [TorchLean.Storage α] (s : EagerSession α) (op : NN.Backend.BackendOp) (handlers : Array (NN.Backend.KernelHandler β)) : IO β

Execute an operation through a handler that matches the selected capsule.

Fail if the chosen provider is not linked into this build. ExecutableKernel certifies dispatch identity; the selected capsule states the numerical evidence.

def Runtime.Autograd.Torch.Internal.EagerSession.executeReferenceOrLibTorch {α β : Type} [TorchLean.Storage α] (s : EagerSession α) (op : NN.Backend.BackendOp) (cpu cuda : IO β) : IO β

Execute an operation implemented by the reference CPU and LibTorch CUDA runtimes.

The selected handler computes tensor values. TorchLean retains tape ownership in either case.

def Runtime.Autograd.Torch.Internal.EagerSession.backendSelections {α : Type} [TorchLean.Storage α] (s : EagerSession α) : IO (Array NN.Backend.AcceptedKernel)

Accepted capsules actually selected by this eager session.